Simultaneously optimizing sampling pattern for joint acceleration of multi-contrast MRI using model-based deep...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Magnetic Resonance Imaging

Background:

• Deep learning is widely used for accelerating Magnetic Resonance Imaging (MRI).
• Joint acceleration of multiple MRI acquisitions is more effective than single-acquisition methods.
• Optimizing sampling patterns enhances MRI acceleration, but this has not been well-explored for joint multi-acquisition acceleration.

Purpose of the Study:

• To develop a model-based deep learning scheme for optimizing sampling patterns in joint acceleration of multi-contrast MRI.
• To enable simultaneous optimization of sampling patterns and reconstruction for multiple MRI contrasts.

Main Methods:

• A physics-guided, model-based deep learning scheme (J-MoDL) was extended to optimize separate sampling patterns for each contrast simultaneously.
• The method was tested on T2-weighted, FLAIR, and T1-weighted images and compared against single-contrast, non-optimized multi-contrast, and common-optimized multi-contrast methods.
• The scheme was also evaluated in a data-driven scenario and tested for generalization on knee MRI datasets.

Main Results:

• The proposed scheme showed superior quantitative and qualitative performance compared to baseline and non-optimized multi-contrast methods.
• Optimizing individual sampling patterns for each contrast outperformed using a single common pattern for all contrasts.
• The optimized patterns exhibited less overlap, and the data-driven approach demonstrated generalizability, with a brain-trained model performing well on knee images.

Conclusions:

• A novel scheme effectively combines sampling optimization and multi-contrast MRI acceleration.
• This integrated approach yields superior performance compared to existing acceleration techniques in MRI.